A multi-institution research team recently reported that veterans with both PTSD and unmet social needs appear to have higher rates of dementia or cognitive impairment years later. The study involved 3,817,970 veterans—and, therefore, millions of data points—so the team relied on support from researchers in the Manning College of Information and Computer Sciences (CICS) at UMass Amherst, who contributed to the processing of the copious amounts of data needed for the large-scale study.

Image

Avijit Mitra, who was a UMass CICS doctoral student at the time and now works as an applied scientist II at Amazon, had developed an NLP (natural language processing, a subfield of AI focused on helping computers understand and generate human language) system while at UMass.

For the PTSD study, Mitra applied that same system to detect health-related social needs (HRSNs)—individual social circumstances that can impact health, such as financial problems, housing instability and food insecurity—and investigate potential connections between HRSNs, PTSD, and mild cognitive impairment, Alzheimer's disease and Alzheimer’s-related dementia.

Traditionally, epidemiological research has depended on claims and coded EHR (electronic health records) fields, said Hong Yu, a CICS adjunct professor who leads the BioNLP Lab at UMass Amherst and UMass Lowell.

EHR narratives such as doctor’s visit notes, discharge summaries and history reports, for example, have often been neglected by epidemiologists due to concerns about bias within the documentation, as well as the difficulty in extracting information reliably, especially on a large scale, Yu noted.

That’s where NLP comes in, she added. This work shows that these systems “can yield associations consistent with those obtained from structured data,” she said. “Although this consistency does not eliminate potential bias, it provides evidence supporting the incorporation of narrative-derived information into epidemiologic research.”

This is the first group to report integrating AI into an epidemiological study, Yu said.

“This is an unprecedented era for computer scientists to engage in multidisciplinary studies,” Mitra said. “With the rapid advancement in ML (machine learning) and rise of LLMs (large language models), it opens paths for computer scientists to collaborate with the researchers in other fields, extending from physics and biology to public health and legal workflows with a special focus on privacy, security and accountability.”

Other UMass CICS researchers who contributed to the study were current doctoral candidate Rumeng Li and Zonghai Yao, who earned his doctorate at UMass Amherst and now is a research assistant professor at UMass Lowell.

The research was published in September 2026 in JAMA Network Open.